分布式图计算框架混合计算模式的研究

发布时间：2018-05-25 03:40

本文选题：分布式 + 图计算　；参考：《小型微型计算机系统》2015年04期

【摘要】：分析了分布式图计算框架的同步和异步计算模式在调度开销和收敛速度上存在的优点与不足.同步计算模式调度开销小,但是收敛较慢;而异步计算模式收敛较快,但调度开销大.基于上述发现,提出一种混合计算模式,能够在分布式环境下有效地结合同步与异步计算模式的优点克服各自不足,以获得最优性能.混合计算模式采用"同步控制流"以降低分布式环境下的调度开销,同时采用"异步数据流"使计算过程使用较新的数据以加快收敛速度.基于多个典型图算法和真实大规模图的评测显示,混合计算模式的性能是原有同步计算模式的1.2倍到2.4倍,计算量平均减少30%;相对于异步计算模式通过减少调度开销,整体性能可以提升至其2.3倍到4.6倍.
[Abstract]:The advantages and disadvantages of the synchronous and asynchronous computing models in the distributed graph computing framework in scheduling overhead and convergence speed are analyzed. The scheduling cost of synchronous computing mode is small, but the convergence is slow, while the asynchronous computing mode converges faster, but the scheduling overhead is large. Based on the above findings, a hybrid computing model is proposed, which can effectively combine the advantages of synchronous and asynchronous computing modes in distributed environments to overcome their respective shortcomings to achieve optimal performance. The hybrid computing model uses "synchronous control flow" to reduce the scheduling overhead in distributed environment, and "asynchronous data stream" to make the computing process use newer data to speed up the convergence. The performance of the hybrid computing model is 1.2 to 2.4 times of that of the original synchronous computing model, and the average amount of computation is reduced by 30 times compared with the asynchronous computing mode, and the scheduling overhead is reduced compared with the asynchronous computing mode. The overall performance can be increased to 2.3 to 4.6 times.
【作者单位】：复旦大学软件学院上海市数据科学重点实验室;上海交通大学软件学院并行与分布式系统研究所;
【基金】：国家自然科学基金青年项目(61003002)资助教育部新教师博士点基金项目(20130073120040)资助
【分类号】：TP338.8

【参考文献】